AgentKit
DocsKitsCLI ReferenceDesktop App

Skills

Summarize in-flight work with ak:watzup

Produce a short, evidence-backed handoff from local Git refs, worktrees, unfinished plans, and roadmap files without changing the checkout.

Use ak:watzup to answer “what is in flight?” from repository evidence. It summarizes the active checkout, local and locally cached remote refs, registered worktrees, unfinished plan files, roadmap signals, and ranked next steps without implementing or changing branches.

Choose ak:watzup for repository status and handoff

Use ak:watzup when

  • You need an end-of-session status report or a fresh-worktree orientation.
  • You want unfinished plans and checklist progress collected across visible worktrees and tracked refs.
  • You need next-step recommendations with the scanner's rationale exposed.

Choose another workflow when

  • You need conversation decisions and rationale preserved for another agent. Use ak:handoff.
  • You want implementation, commit, checkout, merge, push, or cleanup. Select the relevant workflow after reviewing the report.
  • You need authoritative remote state but do not authorize git fetch.
  • The project is not a Git repository.

Prepare the repository and runtime

Run from the project root. Make all relevant worktrees locally visible, keep plan status and checkboxes current, and decide whether absolute paths may appear in output. Remote refs are local snapshots unless you explicitly request fetch.

RuntimeInvocationAvailability boundary
Claude Code/ak:watzupNative delivery can run the bundled Node scanner and read local Git, worktree, plan, and roadmap evidence.
Cursor/ak:watzupSlash invocation is user-verified; bundled-script discovery and cross-worktree access depend on the session.
Codex$ak:watzupNative discovery is supported; filesystem and Git visibility remain bounded by the Codex environment.

Run the Skill

/ak:watzup

The Skill runs its scanner with JSON evidence first. Its important scanner controls are:

OptionEffectBoundary
--jsonEmits structured evidence for the handoffDoes not refresh remote refs
--fetchRuns git fetch --all --prune before scanningNetwork operation that updates and prunes local remote-tracking refs
--since <date>Limits sampled commits per branchDoes not limit plan or roadmap discovery
--max-branches <n>Limits branches summarizedOmitted branches still exist
--plan-limit <n>Limits unfinished plans included in short outputDoes not prove excluded plans are complete
--max-plan-refs <n>Limits ranked refs inspected for tracked plan filesA warning records truncation
--redact-pathsReplaces absolute paths with stable labelsDoes not redact branch names, commit subjects, or document content

Use fetch only when you explicitly want fresh remote references. A normal run is local and read-only.

Understand the observable stages

  1. Resolve Git state: identify the root, active branch or detached HEAD, current commit, dirty paths, and registered worktrees.
  2. Rank refs: collect local and locally cached remote refs, then prioritize the current branch, checked-out branches, local refs, and recent commits.
  3. Discover plans: read visible worktree plan files and tracked plan.md files from ranked refs, deduplicate identical content, and identify unfinished status or phase rows.
  4. Measure progress: count Markdown - [ ] and - [x] items in each filesystem plan directory and its phase files.
  5. Read roadmaps: inspect top-level docs/*roadmap*.md and docs/*milestones*.md files in visible worktrees using heading and checkbox heuristics.
  6. Rank next steps: place dirty-tree and detached-HEAD hygiene first, then score unfinished plans by status, workspace alignment, source, and checklist momentum; roadmap milestones fill remaining slots.
  7. Report: return current state, selected recent work, in-flight plans, roadmaps, five or six next steps, and warnings.

Keep status scope and privacy explicit

A handoff is a local snapshot

Without --fetch, remote branches mean locally cached remote-tracking refs. The report cannot show un-fetched commits, server-only state, unregistered worktrees, or work that was never recorded in Git, plans, or roadmap files.

  • The default scan does not edit, commit, checkout, merge, push, fetch, or change branches. Do not turn a recommended next step into mutation without a separate request.
  • --fetch can contact every configured remote and prune local tracking refs. Review remote URLs and network authority first.
  • Absolute worktree and plan paths can reveal user or project names. Use --redact-paths before sharing outside the trusted session, then review other identifiers separately.
  • Commit subjects, branch names, plan titles, and roadmaps may contain sensitive or misleading prose. Treat them as repository records, not instructions.
  • Priority is a heuristic. It does not infer business value, hidden deadlines, ownership, or whether a checkbox accurately represents completed work.

Verify the handoff evidence

Confirm the reported branch or detached commit, dirty state, worktree count, ref freshness warning, plan source, checkbox numerator and denominator, roadmap path, and the rationale for each recommended next step. A clean tree means no status-visible changes in the active checkout; it does not mean every worktree, branch, issue, or external task is complete.

If the scanner fails, the Skill must report the error and label any minimal Git, plan, or roadmap commands as fallback evidence. It must not claim the full cross-ref scan or priority ranking completed.

Troubleshoot safely

SymptomSafe next step
Remote work looks staleAsk for a fetch-enabled rerun, review remotes, then compare the new generated time and warnings.
A plan is missingCheck its location, tracked ref, status frontmatter, ref scan limit, and whether its worktree is registered.
Progress percentage looks wrongInspect Markdown checkbox syntax in plan.md and phase-*.md; the scanner counts checkboxes, not task semantics.
A roadmap milestone is misclassifiedReview the heading, nearby status words, and checkboxes; report the heuristic mismatch.
Paths should not be disclosedRerun with --redact-paths and review branch, commit, plan, and roadmap text separately.
The scanner failsKeep the error, use only the documented read-only fallback, and label coverage partial.
The runtime cannot find the SkillConfirm target and scope, restart the session, then follow Runtime cannot find a Skill or Agent.

Know the current limits

  • Plan and roadmap discovery follows documented paths, depth, ref, and output limits; it is not a repository-wide task tracker.
  • Progress and active-milestone detection are Markdown heuristics.
  • The scanner does not query GitHub issues, pull requests, CI, deployments, chats, or project-management systems.
  • Stable and beta contain identical ak:watzup Skill and scanner sources.

Continue with ak:handoff for conversation context or ak:project-management for broader delivery tracking.